Two pieces of news from the last few weeks point to the same deeper shift in bio AI. We are moving from models that interpret a dataset to models that treat biology itself as a programmable system, where “cell state” becomes a representation you can query, perturb, and optimize.
MIT covered a new multimodal single cell method that explicitly separates what different assays share from what each assay uniquely measures. Instead of forcing everything into one blended embedding that hides where the signal came from, it builds a shared space plus modality specific spaces. That design choice sounds abstract, but it changes what you can trust. If the shared representation stays stable across RNA, protein, and other readouts, you get a cell state vector that is less dependent on the instrument and more dependent on biology. https://news.mit.edu/2026/ai-help-researchers-see-bigger-picture-cell-biology-0225
The underlying paper describes this as partial information sharing, learned rather than assumed. In practice, this is how you move from “integration for visualization” toward “integration for prediction,” because you can ask which parts of a model’s belief are truly cross modal and which parts are assay artifacts. When people talk about foundation models for cells, this is the kind of representational discipline that makes generalization possible. https://www.nature.com/articles/s43588-025-00948-w
At the same time, Berkeley Lab described OPAL, part of a DOE effort to build self improving foundation models that connect prediction and design to automated experimentation. The framing is blunt: models are not the endpoint, they are components in an autonomous loop that decides what to test next, runs the lab workflow, and learns from the results. If you combine that loop with multimodal cell state representations, biology starts to look like an API with feedback, not a pile of disconnected assays. https://newscenter.lbl.gov/2026/02/02/foundational-ai-models-to-accelerate-biological-discovery/
The headline version is that bio AI is getting more “general.” The more useful version is that it is getting more operational. The winners will be the teams that can standardize how cell state is represented across modalities, then wire that representation into experiment selection and automated execution. When that clicks, the fastest way to do biology will look a lot like the fastest way to do software.
Sources https://news.mit.edu/2026/ai-help-researchers-see-bigger-picture-cell-biology-0225
https://www.nature.com/articles/s43588-025-00948-w
https://newscenter.lbl.gov/2026/02/02/foundational-ai-models-to-accelerate-biological-discovery/